Aucbvax.6888
fa.info-vax
utzoo!decvax!ucbvax!info-vax
Fri Apr 23 07:13:13 1982
instruction timings
>From
[email protected] Fri Apr 23 07:03:35 1982
We tried to get DEC to give us a good idea what timings to
expect when we got our VAX-11/780 in the Medical Center here.
What we got was the line that the precise numbers were company
confidential [!], but they did let us look at (but not copy) a
set of documentation which gave some samples. There *are* some
problems with the simpleminded methods of calculating speed
which work on the lower-end PDP-11's (like a specific factor for
each memory reference which works pretty well for 11/20 through
11/45 but breaks down on the 11/70 because of cache), because
there is a lot of cache, pipelining, and so forth going on. The
figures did have some interesting information, of which what I
remember is something as follows:
o For doing heavy floating-point number crunching, the
780 should be run with the floating point accelerator
if at all possible. As I remember the floating point
times were comparable to a PDP-11/34 [!!] without the
unit (i. e., about 10-40 us) and a bit faster than
the PDP-11/70 with the unit (0.5-1.5 us).
o For purely number-crunching work (if you don't need
decimal or character instructions and if you don't
need a lot of memory for efficient execution), the
780 is perhaps slightly faster than the PDP-11/70.
This does not count the fact that large programs
will be easier to deal with and will run faster on
the 780 because overlays or forks will no longer be
needed, nor does it count the fact that the 780 has
a higher i/o bandwidth.
o DEC claims that the 780's processing power is
roughly that of the IBM-370/155. This seems to be
a reasonable guess from what I've seen.
This was before the 750 (or 730). Benchmarks which we have run
on the 780 and the PDP-11/70 seem to confirm this: for pure
number crunching (assuming the FPA on the VAX and the FPP on the
PDP-11/70), the two machines are roughly equivalent (within the
restrictions given above). I don't have any of the numbers in
front of me at the moment, but they are within about 5-15% of
each other, sometimes one is faster and sometimes the other
(perhaps the vax is usually faster, but you know how hard it is
to get reliable benchmark results).
The 11/750 *is* faster for some types of instructions, e. g.,
if you have the G/H option for the 750 this obviously will be
faster than software emulation for the 780. It will be
interesting to see if this holds when the G/H microcode option
for the 780 comes out later this spring or early summer. The
750 is also faster on shift instructions (ASHL, ROTL) and
perhaps some other instructions because of the superrotator in
the microengine, but I don't know what instructions use its
full possibilities (not having dealt with the 750).
I know very little about the 730 except that it exists and is
"somewhat slower than a 750." I assume that similar differences
will exist between the 730 and the other members of the family
as between the 750 and the 780 -- i. e., some instructions will
take disproportionate amounts of time on that machine compared
to others in the family.
Bruce C. Wright @ Duke University
-----------------------------------------------------------------
gopher://quux.org/ conversion by John Goerzen <
[email protected]>
of
http://communication.ucsd.edu/A-News/
This Usenet Oldnews Archive
article may be copied and distributed freely, provided:
1. There is no money collected for the text(s) of the articles.
2. The following notice remains appended to each copy:
The Usenet Oldnews Archive: Compilation Copyright (C) 1981, 1996
Bruce Jones, Henry Spencer, David Wiseman.
